Invented Predicates to Reduce Knowledge Acquisition

1
Invented Predicates to Reduce Knowledge
Acquisition
2
Aim

• develop ML technique to facilitate KA
• knowledge engineering task of identifying
  intermediate abstractions.
• generalize rule using intermediate abstractions
• base on Duces intra construction and absorption
  operators Muggleton, 1990

3
Ripple Down Rules

• KB added to deal with specific case that wrongly
  classified
• KB evolves over time gradually
• The system organize the structure (not KE or
  expert)
• Any KA validated, not degrade previous knowledge

4
RDR
5
RDR
6
Predicate Invention

• Add a new feature/predicate to enrich the
  language
• Use a new feature to generalize existing
  knowledge

7
Intra-construction

• X B, C, D, E
• X A, B, D, F
• are replaced by
• X B, D, Z
• Z C, E
• Z A, F

8
Absorption

• Y M, N, C, E
• are replaced by
• Y M, N, Z
• while
• Z C, E
• Z A, F

9
Heuristic

• to decide
• applying intra-construction operator (which
  created intermediate concept)
• keeping/deleting intermediate concept
• applying absorption operator (which generalize
  the knowledge based system)

10
(No Transcript)
11
4 simulated expertise

• P uses all the conditions from the rule trace.
  However, for this case we add some more
  refinement rules to cover every case in data set
  correctly.
• G uses one less condition than in the rule trace
  (from P), selected randomly, unless this rule
  causes inconsistency to any seen cases.
• S uses all the conditions from the rule trace
  (from P) but adds another .
• W uses all the conditions from the rule trace
  J4.8 in default mode. This results in some
  errors in the data set. For this simulated
  expert, cases which have errors are removed from
  the data set before building the RDR KBS.

12
KA sessions for Car data set
13
KA sessions for Nursery data set
14
Conclusion

• Reduce KA effort up to 49
• Propose a learning system that supports knowledge
  acquisition without increase the task of the
  human expert
• These results confirm the value of heuristic
  classification with its intermediate structure
  as suggested by Clancey 1985.